Chapter 8 Small Units inside Large Units

8.1
8.1.1

Experimental units bigger than observational units

The context

Example 8.1 (Example 1.2 continued: Calf-feeding) Four feed treatments are compared on 80 calves. The calves are kept in eight pens, with 10 calves per pen. Each feed is applied to two whole pens, with calves feeding ad lib. Even though every calf is weighed individually, differences between feeds should be compared with differences between pens, not with differences between calves. So should we say that the replication is 20 or 2? Example 1.8 is similar: treatments are applied to whole classes but individual children are measured. So is one version of Example 1.11: pullets are fed but eggs are weighed. So are Examples 7.7 and 7.8: treatments are applied to whole doctors practices or to villages, but it is individual people that are measured. In general, suppose that there are m experimental units, each of which consists of k observational units, and that there are t treatments, each of which is applied to s experimental units, so that st = m. 8.1.2 Construction and randomization

This is just like the completely randomized design in Section 2.1, using the experimental units instead of the plots. 8.1.3 Model and strata

For simplicity, I use the language of Example 8.1. If pens give xed effects then we cannot estimate anything about treatments, so we assume that pens give random 129

130

Chapter 8. Small Units inside Large Units

effects. This gives a model like the random-effects model in Sections 4.4 and 4.6: E(Y ) = T () for each calf , and 2 cov(Y , Y ) = 1 2 2 2 if = if and are different calves in the same pen if and are in different pens.

Usually 1 > 2 > 0 unless there is so little food that there is competition among calves in the same pen. Put VP = {vectors in V which take a constant value on each pen} and WP = VP V0 . From Section 4.6, we see that the eigenspaces of Cov(Y) are with dimensions and eigenvalues where 0 = 2 (1 1 ) + 2 k(1 2 ) + 2 bk2 1 = 2 (1 1 ) + 2 k(1 2 ) 2 = 2 (1 1 ). If 1 > 2 then 1 > 2 ; if 2 > 0 then 0 > 1 . 8.1.4 Analysis Here I show how to build up the analysis-of-variance table in stages, so that it does not have to be memorized. The rst step is to write out the list of strata and their degrees of freedom in a table, with one row for each stratum. This table is called the null analysis of variance. It is shown in Table 8.1. The next two steps can be done in either order. The second step can be done before treatments are thought about. Simply take the rows of the null analysis-ofvariance table, and show the calculations associated with each one. See Table 8.2. The rst column contains sums of squares. As usual, the sum of squares for the mean is sum2 /N . The sum of squares for pens, SS(pens), is equal to CSS(pens) SS(mean), where m sum2 pen=i . CSS(pens) = k i=1 V0 1 0 WP m1 1
VP m(k 1) 2 ,

8.1. Experimental units bigger than observational units Stratum mean pens calves Total df 1 m1 m(k 1) mk = N SS sum2 /N i sum2 pen=i /k SS(mean) y2 CSS(pens) y2 EMS 0 1 2

131

Table 8.1: Null analysis of variance The total sum of squares is

Table 8.2: Calculations ignoring treatments

y2 ,

and the sum of squares for calves is obtained by subtraction, so that it is equal to y2 CSS(pens). The second column contains the expected mean squares if there are no treatments. These are just the appropriate eigenvalues of Cov(Y). The third step is to take the null analysis of variance and expand it to show where treatments lie. This gives the skeleton analysis of variance, shown in Table 8.3. An extra column for source is inserted after the stratum column, and the degrees-offreedom column is split into two. If there is any stratum that contains no treatment subspace (except for V0 ), then its name is copied into the source column and its degrees of freedom are written on the right-hand side of the df column. This rule gives the rows labelled mean and calves in Table 8.3. The next lemma shows why there are no treatment subspaces in the calves stratum. Lemma 8.1 If treatments are applied to whole pens then WT WP . Proof Let v be in VT . If and are in the same pen then T () = T () so v = v because v VT . Hence v VP . Therefore VT VP , so WT = VT V0 VP V0 = WP . Since the pens stratum contains WT , which is the subspace for differences between feeds, the row labelled pens in the null analysis-of-variance table has to be split into three, with names shown in the source column. The rst is WT , labelled feed, whose degrees of freedom are shown on the left-hand side of the df column. The third is labelled total; it is really a copy of the pens row in Table 8.1 so its degrees of freedom are shown on the right of the df column. In . Its degrees between is the source labelled residual, which is the space WP WT of freedom, shown on the left of the df column, are calculated by subtraction as (m 1) (t 1) = m t . The fourth step is to combine Tables 8.2 and 8.3 and insert extra information to give Table 8.4, which is the full analysis of variance. The sums of squares for the

132

Chapter 8. Small Units inside Large Units Stratum mean pens source degrees of freedom mean 1 feed t 1 residual m t total m1 calves m(k 1) N

calves Total

Table 8.3: Skeleton analysis of variance mean and calves strata are copied from Table 8.2 to Table 8.4. In the pens stratum, the sum of squares is copied from Table 8.2 to the total line of the pens stratum in Table 8.4. The sum of squares for treatments is calculated in the usual way as SS(feed) = sum2 sum2 T =i . sk N i=1
t

The sum of squares for residual is obtained by subtraction: SS(residual) = SS(total) SS(feed). If there is space, it is useful to split the sum of squares column into two just like the df column. Expected mean squares are copied from Table 8.2 to Table 8.4, being put in the right-hand side of the EMS column. Thus 0 is written for the mean, 2 for calves, and 1 for both of the non-total sources in the pens stratum. To these stratum variances must be added sums of squares of expected values. As in Section 4.5, E(PV0 Y) = 0 and E(PWT Y) = T , so Theorem 2.10(ii) shows that EMS(mean) = 0 2 + 0 , T 2 EMS(feed) = + 1 , t 1 EMS(residual) = 1 and EMS(calves) = 2 . The nal column shows what is the appropriate variance ratio, if any. There is no independent estimate of 0 , so there is no test for whether 0 is zero. If T = 0 then EMS(feed) = EMS(residual) so the appropriate variance ratio to test for differences between feeds is MS(feed) . MS(residual) 8.1.5 Hypothesis testing

Test for difference between feeds by using the variance ratio MS(feed) . MS(residual)

Stratum 1 t 1 1 sum2 sum2 T =i sk N T 2 + 1 t 1 sum2 N 0

2

source + 0 MS(feed) MS(residual)

df

SS

EMS

VR

mean

mean

pens

feed

residual . . . . . . . . . . . . . . by subtraction . . . . . . . . . . . . . . m1 m(k 1) N y2 y2 Pi2 k Pi2 sum2 k N 2

total

calves

calves

Total

8.1. Experimental units bigger than observational units

Table 8.4: Full analysis of variance: here Pi denotes the total of the data on the i-th pen

133

134

Chapter 8. Small Units inside Large Units

The appropriate numbers of degrees of freedom for the F-test are t 1 and m t . If s = 1 then m = t and so there is no residual in the pens stratum and so no test can be done. In this case people sometimes wrongly compare MS(feed) to MS(calves). If 1 > 2 then 1 > 2 so this test is likely to detect differences that do not really exist. This is known as false replication or pseudo-replication, because the replication of calves within the pens is falsely taken as a substitute for the replication of pens within treatments. Probably the single most common mistake in designing experiments is to allocate treatments to large units with s = 1. This is the mistake in the ladybirds experiment in Example 1.1. Even when s = 1, some people mistakenly combine the sum of squares for residual in the pens stratum with the calves sum of squares, to give a source whose expected mean square is equal to (m t )1 + m(k 1)2 . N t This is smaller than 1 if 1 > 2 , so again there is spurious precision. There are also more degrees of freedom than there should be, which leads to spurious power. Note that increasing k does not increase the number of degrees of freedom for residual. 8.1.6 Decreasing variance

Theorem 2.10(iv) shows that the variance of the estimator of a simple treatment difference i j is equal to 2 2 (1 1 ) 1 = 2 + (1 2 ) , sk s k which is estimated by (2/sk) MS(residual). To decrease the variance, it is more fruitful to increase s than to increase k, even though the experimenter may nd it simpler to increase k. Example 8.1 revisited (Calf-feeding) Plausible values for the two correlations are 1 = 0.3 and 2 = 0. Then the variance of the estimator of a difference is equal to 2 2 0.7 0.3 + . sk s

Some combinations of values of s and k give the following coefcients of 22 . k = 7 k = 10 k = 15 s = 2 0.200 0.185 0.173 s = 3 0.133 0.123 0.116 The original design has variance 22 0.185. Increasing the number of calves per pen from 10 to 15 only decreases the variance to 22 0.173, even though 50%

8.2. Treatment factors in different strata

135

more feed is used. On the other hand, replacing 8 pens of 10 calves by 12 pens of 7 calves decreases the variance substantially to 22 0.133 while using almost the same number of calves and almost the same amount of feed. However, the extra four pens may not be available, or they may be expensive to set up. It is almost always easier for the experimenter to randomize treatments to a few large units than to many smaller units. Part of the statisticians job is to explain the loss of precision and power that will result from this.

8.2

Treatment factors in different strata

Example 8.1 revisited (Calf-feeding) Suppose that the four feeds consist of all combinations of two types of hay, which is put directly into the pen, with two types of cake, which is fed to calves individually. If all calves in the same pen have the same type of cake then the design and analysis are just as in Section 8.1, except that the treatments line in the analysis of variance is split into three, giving the skeleton analysis of variance in Table 8.5. Stratum mean pens source mean hay cake hay cake residual total calves degrees of freedom 1 1 1 1 4 7 72 80

calves Total

Table 8.5: Skeleton analysis of variance when types of hay and types of cake are both applied to whole pens It might be better to give ve calves in each pen one type of cake and the other ve calves the other type of cake. In general, suppose that there are m large units, each of which consists of k small units; and that there are nH levels of factor H , each of which is applied to rH large units (so that nH rH = m); and that there are nC levels of treatment factor C, each of which is applied to rC small units per large unit (so that nC rC = k). Now the number of treatments is nH nC and they all have replication rH rC . Thus N = mk = nH nC rH rC . The construction and randomization are as follows. (i) Apply levels of H to large units just like a completely randomized design. (ii) Within each large unit independently, apply levels of C just like a completely randomized design.

136

Chapter 8. Small Units inside Large Units

Let L be the factor for large units, so that L() is the large unit which contains small unit . Dene subspaces VL and WL like the subspaces VP and WP in Section 8.1. Theorem 8.2 If levels of H are applied to large units with equal replication and levels of C are applied to small units with equal replication within each large unit, then (i) WH WL ;
; (ii) WC VL . (iii) WH C VL

Proof

(i) Use Lemma 8.1.

(ii) and (iii) Let be any large unit. Suppose that H has level h on . Let v be . The r large units on any vector in WC or WH C . Then v VT and v VH H which H has level h all have the same treatments, so they all have the same values of v, although probably in different orders. Therefore

v = rH v,

(8.1)

where is the union of the large units on which H takes level h. Let w be the vector whose entry on plot is equal to 1 0 if otherwise,

and dene the vector w analogously. Equation (8.1) can be rewritten as v w = rH v w . Now, w is in VH and v is orthogonal to VH , so v w = 0 so v w = 0. This is true for all large units , and VL is spanned by the vectors w as runs over the large units, so v is orthogonal to VL . The analysis-of-variance table is built up as in Section 8.1. The null analysis of variance is shown in Table 8.6 and the calculations ignoring treatments in Table 8.7. Theorem 8.2 shows us how to expand Table 8.6 to the skeleton analysis of variance in Table 8.8. The row for the large units stratum must be split to give a line for H , a residual line and a total line. The total line is the same as in the null analysis of variance, and the residual is the difference between total and H . The row for the small units stratum must be split even further: it has to show two treatment lines, the main effect of C and the H -by-C interaction; the residual is obtained by subtracting both of these from the total. Note that we now have two separate lines called residual, one in the large units stratum and one in the small units stratum.

8.2. Treatment factors in different strata Stratum mean large units df 1 m1 SS sum2 /N i sum2 L=i /k SS(mean) y2 CSS(large units) y2 EMS 0 1 2

137

small units m(k 1) Total mk = N

Table 8.6: Null analysis of variance Stratum mean large units source mean H residual total small units C H C

Table 8.7: Calculations ignoring treatments degrees of freedom 1 nH 1 m nH m1 nC 1 (nH 1)(nC 1)

residual m(k 1) nH (nC 1) total Total Table 8.8: Skeleton analysis of variance Where necessary, I distinguish these by writing, for example, large units residual in full. Tables 8.7 and 8.8 are combined to give the full analysis of variance in Table 8.9. Here 0 = PV0 , H = PWH , C = PWC , HC = PWH C , CSS(L) = sum2 kL=l , l =1
m

m(k 1) N

SS(H ) =

sum2 sum2 H =i , N i=1 rH k

nH

SS(C) =

j=1

nC

2 sumC =j

rC m

sum2 N

138 and

Chapter 8. Small Units inside Large Units

SS(H C) =

nH nC

i=1 j=1

sum2 H =i,C= j rH rC

sum2 SS(H ) SS(C). N

Now differences between levels of H are assessed against the variability of large units. Testing uses MS(H ) . MS(large units residual) The standard error of a difference between levels of H is 2 rH k MS(large units residual),

and the variance of the estimator of such a difference is 21 /rH k. The main effect of C and the H -by-C interaction are assessed against the variability of small units within large units. The tests use the variance ratios MS(C) MS(small units residual) and MS(H C) MS(small units residual)

respectively. The variance of the estimator of the difference between two levels of C is 22 /rC m, so the standard error of a such a difference is 2 MS(small units residual). rC m Standard errors of differences between indiviual treatments require a little more care. For i = 1, . . . , nH and j = 1, . . . , nC , let ui j be the vector in VT whose coordinate on plot is equal to 1 if H () = i and C() = j 0 otherwise. Similarly, let vi be the vector whose coordinate on is equal to 1 0 if H () = i otherwise

and w j the vector whose coordinate on is equal to 1 0 if C() = j otherwise.

Fix levels i, i of H and j, j of C, with (i, j) different from (i , j ). Put u = (ui j ui j )/r, where r = rH rC . Then u V0 . Let v be the projection of u onto

Stratum 1 SS(H ) 1 H 2 + 1 nH 1 MS(H ) MS(large units residual) sum2 N 0

2

source + 0

df

SS

EMS

VR

mean

mean

large units

nH 1

residual m1 SS(C) SS(H C) HC 2 2 dHC 2 C 2 + 2 nC 1 CSS(L) sum2 N

. . . . . . . . . . . by subtraction . . . . . . . . . . .

total

small units

nC 1

MS(C) MS(small units residual) MS(H C) MS(small units residual)

H C

dHC

8.2. Treatment factors in different strata

residual m(k 1) N y2 y2 CSS(L)

. . . . . . . . . . . by subtraction . . . . . . . . . . .

total

Total

Table 8.9: Full analysis of variance: here dHC = (nH 1)(nC 1) 139

140

Chapter 8. Small Units inside Large Units

VH . Theorem 2.3(vii) shows that v = (vi vi )/rnC , which is in WH , so v = PWH u. Similarly, if w = PWC u then w = (w j w j )/rnH . Therefore u = v + w + x, (8.2)

where x WH C . The vectors on the right-hand side of Equation (8.2) are mutually orthogonal, and so x 2 = u 2 v 2 w 2. Now, u 2 = 2/r; v 2 is equal to 2/rnC if i = i and to zero otherwise; and w is equal to 2/rnH if j = j and to zero otherwise. Therefore 2(nH nC nH nC )/N if i = i and j = j 2 x = 2nC (nH 1)/N if i = i and j = j 2nH (nC 1)/N if i = i and j = j .
2

Parts (iii) and (iv) of Theorem 2.10 show that v Y is the best linear unbiased estimator of v and that it has variance v 2 1 . Similarly, w Y is the best linear unbiased estimator of w and it has variance w 2 2 , and x Y is the best linear unbiased estimator of x and it has variance x 2 2 . Hence u Y is the best linear unbiased estimator of u , which is equal to i j i j . Theorem 2.10(v) shows that three estimators are uncorrelated, and so Var(u Y) = v 2 1 + w 2 2 + x 2 2 . If i = i while j = j then Var(u Y) = 2nC (nH 1) 2 + 2 rnH N 2nC = [1 + (nH 1)] 2 N 2nC nH = 2 N 2 = 2 . r

If i = i but j = j then Var(u Y) = 21 2nH (nC 1)2 + rnC N 2 = [nH 1 + nH (nC 1)2 ] N 2 = [1 + (nC 1)2 ] . rnC

8.2. Treatment factors in different strata If i = i and j = j then Var(u Y) = 21 22 2(nH nC nH nC )2 + + rnC rnH N 2 = [nH 1 + nH (nC 1)2 ] N 2 = [1 + (nC 1)2 ] . rnC

141

Thus the standard error of a difference is equal to 2 MS(small units residual) r if the levels of H are the same; otherwise it is equal to 2 [MS(large units residual) + (nC 1) MS(small units residual)]. rnC Since we expect 2 to be smaller than 1 , this design gives more precise estimates of differences between levels of C than the design in Section 8.1. It also gives more precise estimates of differences between any two levels of H C. There is increased power for testing the main effect of C and the H -by-C interaction, for not only is 2 smaller than 1 but the number of residual degrees of freedom in the small units stratum is substantially more than the number of residual degrees of freedom in the large units stratum in the design in Section 8.1. There is even a slight increase in power for testing the main effect of H , because the number of residual degrees of freedom in the large units stratum increases from m nH nC to m nH . Example 8.1 revisited (Calf-feeding) If each type of hay is applied to four whole pens and each type of cake is given to ve calves per pen, then we obtain the skeleton analysis of variance in Table 8.10. The improved precision for cake and for hay cake is immediately evident. To see the improved power for testing for the main effect of hay, we argue as we did in Sections 2.12 and 4.8. There is one degree of freedom for hay, and an F-test with one degree of freedom for the numerator is equivalent to a two-sided t-test on the square root of the statistic. The vector H has just two distinct entries, one for each type of hay. Let be the modulus of their difference. The design summarized in Table 8.5 has 4 residual degrees of freedom in the pens stratum. If we test at the 5% signicance level then we need the 0.975 point of the t-distribution on 4 degrees of freedom, which is 2.776. The 0.90 point is 1.553. Thus the argument of Section 2.12 shows that to have probability at least 0.9 of detecting that the main effect of hay is non-zero we need > (2.776 + 1.533) 2 1 = 4.309 40 1 . 20

142

Chapter 8. Small Units inside Large Units Stratum mean pens source degrees of freedom mean 1 hay 1 residual 6 total 7 cake 1 hay cake 1 residual 70 total 72 80

calves

Total

Table 8.10: Skeleton analysis of variance when types of hay are applied to whole pens and each type of cake is given to ve calves per pen On the other hand, the design summarized in Table 8.10 has 6 residual degrees of freedom in the pens stratum. The 0.975 and 0.90 points of the t-distribution on 6 degrees of freedom are 2.447 and 1.440 respectively. In order to have probability at least 0.9 of detecting that the main effect of hay is non-zero, we need > (2.447 + 1.440) 2 1 = 3.887 40 1 . 20

Thus the second design can detect a difference between the types of hay which is 10% smaller than the difference detectable by the rst design.

8.3

Split-plot designs

8.3.1 Blocking the large units In the situtation described in either of Sections 8.18.2, the large units may themselves be grouped into blocks, as in Chapter 4. Then it is convenient to think of the large units as small blocks and the blocks as large blocks. Thus we have b large blocks each of which contains s small blocks each of which contains k plots. Such large and small blocks often occur naturally. Example 8.2 (Example 1.5 continued: Rye-grass) Here the large blocks are the elds and the small blocks are the strips. Thus b = 2, s = 3 and k = 4. Example 8.3 (Animal-breeding) In animal breeding it is typical to mate each sire (male parent) with several dams (female parents), each of which may then produce several offspring. Then the large blocks are the sires, the small blocks are the dams, and the plots are the offspring. Confusingly, the small blocks are sometimes called whole plots or main plots while the plots are called subplots. This is a historical accident based on a famous experiment at Rothamsted Experimental Station.

8.3. Split-plot designs

143

Example 8.4 (Example 5.11 continued: Park Grass) After the Park Grass experiment had been running for many decades, the soil became over-acidic. In 1903 each plot was split into two subplots. Thereafter, lime was applied to one subplot per plot in every fourth year. Although the subplot to be limed was not chosen randomly in the way described below, as far as I know this is the origin of the idea of splitting plots into subplots for the application of levels of a later treatment factor. For simplicity, I shall describe only the classic split-plot design. This is like the second design in Section 8.2 except that the large units (small blocks) are grouped into b large blocks of size s; each level of H is applied to one small block per large block (so nH = s and rH = b); each level of C is applied to one plot per small block (so nC = k and rC = 1). Example 8.2 is a classic split-plot design. 8.3.2 Construction and randomization

(i) Apply levels of H to small blocks just as in a complete-block design. (ii) Within each small block independently, apply levels of C just as in a completely randomized design. 8.3.3 Model and strata

We assume that large blocks and small blocks both give random effects. Thus E(Y ) = H ()C() for each plot , and 2 2 1 cov(Y , Y ) = 2 2 3 2

if = if and are different plots in the same small block if and are in different small blocks in the same large block if and are in different large blocks.

(8.3)

Usually 1 > 2 > 3 . Dene VB as in Chapter 4 and put WB = VB V0 , so that dim VB = b and dim WB = b 1. Further, put VS = {vectors in V which take a constant value on each small block} .

144

Chapter 8. Small Units inside Large Units

Then dim VS = bs. The argument in the proof of Lemma 8.1 shows that VB VS , so it . Then dim W = dim V dim V = bs b = b(s 1). is natural to put WS = VS VB B S S Finally, dim VS = N bs = bsk bs = bs(k 1). The proof of the following proposition is an extended version of the argument that nds the eigenspaces and eigenvectors of Cov(Y) in Section 4.6. Proposition 8.3 Under assumption (8.3), the eigenspaces of Cov(Y) are W0 , WB , WS and VS , with eigenvalues 0 , B , S and respectively, where 0 B S = = = = 2 [(1 1 ) + k(1 2 ) + ks(2 3 ) + bks3 ] 2 [(1 1 ) + k(1 2 ) + ks(2 3 )] 2 [(1 1 ) + k(1 2 )] 2 (1 1 ).

8.3.4 Analysis Stratum W0 WB WS VS mean large blocks small blocks plots Total df 1 b1 b(s 1) bs(k 1) bsk SS sum2 /N i sum2 B=i /sk SS(mean) j sum2 S= j /k CSS(B) y2 CSS(S) y2 EMS 0 B S

Table 8.11: Null analysis of variance Table 8.12: Calculations ignoring treatments for the classic split-plot design for the classic split-plot design Proposition 8.3 shows that the null analysis of variance is as shown in Table 8.11. The calculations ignoring treatments are shown in Table 8.12, where S is the factor for small blocks, CSS(B) = CSS(large blocks) = and CSS(S) = CSS(small blocks) =
j=1 bs

sum2 B=i , sk i=1

sum2 S= j k

To work out the skeleton analysis of variance, we have to decide which stratum contains each treatment subspace. Every treatment occurs once in each block, so Theorem 4.1 shows that WT is orthogonal to VB . The argument of Lemma 8.1 shows = W . Theorem 8.2 shows that WH VS . Since WH WT , we obtain WH VS VB S that WC and WH C are both contained in VS . Hence we obtain the skeleton analysis of variance in Table 8.13. Putting all the parts together gives the full analysis of variance in Table 8.14.

8.4. The split-plot principle Stratum mean large blocks small blocks source mean large blocks H residual total plots C H C residual total Total k1 (s 1)(k 1) (b 1)s(k 1) bs(k 1) bsk s1 (b 1)(s 1) b(s 1) degrees of freedom 1 b1

145

Table 8.13: Skeleton analysis of variance for the classic split-plot design 8.3.5 Evaluation Just as in Section 8.2, C and H C are estimated more precisely than H , and with more power. So why would anyone deliberately choose a split-plot design? There are three good reasons. (i) Treatment factor H is already applied to small blocks in a long-term experiment. At a later stage, it is decided to include treatment factor C, and it is practicable to apply levels of C to smaller units than the small blocks. This is what happened in Example 8.4. (ii) It is not feasible to apply levels of H to such small units as is feasible for levels of C. This is the case in Example 8.2. Unless sowing is done by hand, varieties typically must be sown on relatively large areas while levels of other agricultural treatment factors, such as fertilizer, fungicide or pesticide, can be applied to smaller areas. (iii) The differences between levels of H are already known and the main purpose of the experiment is to investigate C and the H -by-C interaction. If none of those three conditions applies, think twice before recommending a split-plot design.

8.4

The split-plot principle

The split-plot idea is so simple to describe and to implement that there is a temptation to overuse it. Suppose that there are treatment factors F , G, H , . . . . Some scientists like to design experiments with repeated splitting as follows.

Stratum mean Chapter 8. Small Units inside Large Units large blocks large blocks H residual total plots H C residual total Total 146 C k1 (s 1)(k 1) (b 1)s(k 1) bs(k 1) bsk (b 1)(s 1) b(s 1) SS(C) SS(H C) by subtraction s1 b1 SS(H ) by subtraction small blocks mean 1

source

df

SS sum2 bsk CSS(B) SS(mean)

EMS

VR

+ 0

H 2 + S s1

MS(H ) MS(small blocks residual)

S CSS(S) CSS(B)

C 2 + k1

MS(C) MS(plots residual)

HC 2 + (s 1)(k 1)

MS(H C) MS(plots residual)

2 CSS(S) y 2 y

Table 8.14: Full analysis of variance for the classic split-plot design

8.4. The split-plot principle

147

(i) Construct a complete-block design for F , or an equireplicate completely randomized design for F . (ii) Split each plot into nG subplots and apply levels of G randomly to subplots in each plot, independently within each plot. (iii) Split each subplot into nH sub-subplots and apply levels of H randomly to sub-subplots within each subplot, independently within each subplot. (iv) And so on. This is simple to understand, to construct and to randomize. It makes the experiment easy to conduct. There may be practical reasons why levels of F have to be applied to larger units than levels of G, and so on. If not, this is a bad design, because F has higher variance and fewer residual degrees of freedom than G and F G, which in turn have higher variance and fewer residual degrees of freedom than H , F H , G H and F G H , and so on. This matters less if the main purpose of the experiment is to test for interactions. Example 8.5 (Insecticides on grasshoppers) An experiment was conducted on a prairie in Western Canada to nd out if insecticides used to control grasshoppers affected the weight of young chicks of ring-necked pheasants, either by affecting the grass around the chicks or by affecting the grasshoppers eaten by the chicks. Three insecticides were used, at low and high doses. The low dose was the highest dose recommended by the department of agriculture; the high dose was four times as much as the recommended dose, to assess the effects of mistakes. The experimental procedure took place in each of three consecutive weeks. On the rst day of each week a number of newly hatched female pheasant chicks were placed in a brooder pen. On the third day, the chicks were randomly divided into twelve groups of six chicks each. Each chick was given an identication tape and weighed. On the fourth day, a portion of the eld was divided into three strips, each of which was divided into two swathes. The two swathes within each strip were sprayed with the two doses of the same insecticide. Two pens were erected on each strip, and one group of pheasant chicks was put into each pen. For the next 48 hours, the chicks were fed with grasshoppers which had been collected locally. Half the grasshoppers were anaesthestized and sprayed with insecticide; the other half were also anaesthestized and handled in every way like the rst half except that they were not sprayed. All grasshoppers were frozen. The experimenters maintained a supply of frozen grasshoppers to each pen, putting them on small platforms so that they would not absorb further insecticide from the grass. In each swath, one pen had unsprayed grasshoppers while the other had grasshoppers sprayed by the insecticide which had been applied to that swath. At the end of the 48 hours, the chicks were weighed again indiviudally. Table 8.15 shows the skeleton analysis of variance.

148

Chapter 8. Small Units inside Large Units Stratum mean weeks strips source mean weeks insecticide residual total dose insecticide dose residual total food insecticide food dose food insecticide dose food residual total chicks degrees of freedom 1 2 2 4 6 1 2 6 9 1 2 1 2 12 18 180 216

swathes

pens

chicks Total

Table 8.15: Skeleton analysis of variance for the grasshopper experiment

Questions for Discussion

8.1 Read the rst ve pages of the paper Effects of Rodeo R and Garlon R 3A on nontarget wetland species in central Washington, written by Susan C. Gardner and Christian E. Grue and published in the journal Environmental Toxicology and Chemistry in 1996 (volume 15, pages 441451). (a) Describe the experimental design used, and comment on it. (b) Suggest a better design. (c) Comment on the statistical analysis used. 8.2 The following extract is taken from Manual of Crop Experimentation by Pearce, Clarke, Dyke and Kempson. Calculate the analysis-of-variance table, the means for the two main effects, the standard error of a difference between two varieties, and the standard error of the difference between two cutting schemes. An example from Statistical Methods by G. W. Snedecor and W. G. Cochran (1967, p. 370) used three varieties of alfalfa (lucerne) on the main-plots in six randomized blocks. Each of the main plots was divided into four sub-plots, the sub-plots treatments being four cutting schemes. All subplots were cut twice: the second cut took place on 27th July. (The data of the rst are not given.) Some of the plots received a further cut as follows: B, 1st September; C, 20th September; D, 7th October, but A

8.4. The split-plot principle was not cut further. The yields in tons per acre for the following year are set out below. Variety Ladak Cutting A B C D A B C D A B C D Block I II III IV V VI 2.17 1.88 1.62 2.34 1.58 1.66 1.58 1.26 1.22 1.59 1.25 0.94 2.29 1.60 1.67 1.91 1.39 1.12 2.23 2.01 1.82 2.10 1.66 1.10 2.33 2.01 1.70 1.78 1.42 1.35 1.38 1.30 1.85 1.09 1.13 1.06 1.86 1.70 1.81 1.54 1.67 0.88 2.27 1.81 2.01 1.40 1.31 1.06 1.75 1.95 2.13 1.78 1.31 1.30 1.52 1.47 1.80 1.37 1.01 1.31 1.55 1.61 1.82 1.56 1.23 1.13 1.56 1.72 1.99 1.55 1.51 1.33

149

Cossack

Ranger

8.3 On Monday 3rd March 1997 a surgeon sent a message to Allstat, the email list for British statisticians, asking their advice. He said that he was going to use 9 animals in an experiment to compare three methods of grafting skin. He would use 3 animals for each method. After the graft was complete he would take a sample of new skin from each animal. He would cut each sample into 20 (tiny!) pieces and use a precision instrument to measure the thickness of each piece. He wanted to know how to analyse the data. (a) What are the observational units, and how many are there? (b) What are the experimental units, and how many are there? (c) What are the treatments, and how many are there? (d) Construct the skeleton analysis-of-variance table. 8.4 A cattle breeder wants to nd a way of protecting his cattle against a particular stomach disease. He wants to compare the effects of: S: N: spraying the grass in the paddock with a special chemical no spray

He also wants to compare the effect of +: injecting each animal with a special vaccine : no injection

150

Chapter 8. Small Units inside Large Units

He has several paddocks. Each paddock contains 20 animals, labelled 1, . . . , 20 with ear-tags. He wants to apply the treatments once. A month later he will assess the amount of stomach disease in each animal by counting the number of a certain type of bacterium in a sample from the stomach contents. The logarithm of this number will be analysed. He wants to use two paddocks to nd out the effect of the spray, using S on one paddock and N on the other. He wants to use another two paddocks to nd out the effect of the injection, using + on one paddock and on the other. He asks you: Since N and are both no treatment, can I just use three paddocks, one for S, one for + and one for no treatment? How do you answer?